Privacy-first telemetry and monitoring for AI agents

Know when your agents stall, loop or overspend, without sending us a single prompt.

Every run and the steps inside it, every model and tool call with its tokens, latency and cost, and the alert when something goes wrong. Prompts and completions stay out unless an organization owner turns them on for a project.

pip install telmoni

A recording of the console, coming soon

01

What is Telmoni?

An open source, privacy-first telemetry and monitoring platform for AI agents: SDKs in four languages record an agent's runs, ingest keeps them as metadata, and a console reads them as traces, sessions and users.

  • Traces Every run and the steps inside it, each model and tool call with its tokens, latency and failures, over OpenTelemetry's own protocol.
  • Sessions and users Runs grouped by conversation and by the end user they served, under ids hashed with your project's own key.
  • Notifications What happens in a project, in the console's feed, on Slack and through signed webhooks.
  • The console agent Ask about a project without leaving the console, each answer citing the pages it read; read-only, on a model you choose.
  • Coding agents A plugin each for Claude Code, Codex, Cursor and OpenCode records their sessions, and the skill teaches any agent to drive the CLI and the API.
  • Self-host One Rust binary and one console, Apache-2.0, on Docker Compose or Kubernetes; you bring the identity provider, the mail transport and the model.

Next, built in the open: alerts over a window with a warning and an alert threshold, to Slack, webhooks and PagerDuty; what each call cost; and content an owner chooses to keep.

Read the docs →

02

Built for privacy first

  • Prompts, completions and tool payloads come off a span before it leaves your process, and ingest drops any that arrive: what is kept is metadata — timings, statuses, counts, and model and tool names.
  • End users' ids are hashed under each project's own key, every field is documented in the field inventory, and every change to your organization is in its audit log.
Read the field inventory →

03

Work from a terminal, or send an agent

  • The console agent Ask what a project has been doing, who is in it and what changed, each answer with its citations; ⌘J in the console. Read the documentation
  • The agent skill Teaches Claude Code, Codex, Cursor and any agent that reads SKILL.md to drive the CLI, call the API and self-host. Install the skill
  • The CLI Device sign-in for a terminal and a token for CI; organizations, status and configuration from a shell. Install the CLI
  • The HTTP API The read-only API under /v1, with a project API key, and signed webhooks for what happens. Read the reference

04

Questions and answers

What runs today?

The foundation — organizations and roles, API keys, the audit log, notifications, the console agent and the CLI — and the first of the telemetry: the SDKs record an agent's runs, ingest keeps them as metadata, and the console reads them as traces, sessions and users. Next: alert rules and PagerDuty, what each call cost, and content an owner chooses to keep.

What does it store?

Metadata: timings, statuses, counts, and model and tool names. Prompts, completions and tool payloads are not sent by the SDKs and are dropped at the door. A content mode that keeps them, for an organization owner to turn on per project, is planned; today the only mode is off.

Can I self-host it?

Yes, the whole platform: Docker Compose on one machine, or the Helm chart on Kubernetes. You bring the identity provider, the mail transport and the model.

How do I get started?

Sign up, create a project in the organization your first sign-in makes, make an API key and install an SDK; or install the agent skill and let a coding agent do it.